Xianqi Wang, Hao Yang, Gangwei Xu, Junda Cheng, Min Lin, Yong Deng, Jinliang Zang, Yurui Chen, Xin Yang
Huazhong University of Science and Technology, Autel Robotics, Optics Valley Laboratory
Used title: StereoGen: High-quality Stereo Image Generation from a Single Image
- 09/16/2025: Update MfS35K to Hugging Face.
- 09/16/2025: Update MfS35K to Baidu Netdisk.
- 07/29/2025: Update the fine-tuning code for SDv2I.
- 07/14/2025: Update the generation code to improve the quality of the right image edges.
conda create -n zerostereo python=3.12
conda activate zerostereo
pip install tqdm numpy wandb opt_einsum hydra-core
pip install scipy torch torchvision diffusers transformers opencv-python matplotlib
pip install xformers accelerate scikit-image timm==0.5.4
Data for fine-tuning:
Data for generation:
The filepath format should be consistent with the filelist.
Data for training:
Data for evaluation:
| Model | Link |
|---|---|
| SDv2I | Download 🤗 |
| StereoGen | Download 🤗 |
| Zero-RAFT-Stereo | Download 🤗 |
| Zero-IGEV-Stereo | Download 🤗 |
The link to the original SDv2I is invalid. Please use the copy from others, like Download 🤗.
To fine-tune SDv2I, run:
accelerate launch train_stereogen.py
To generate MfS35K, run:
accelerate launch generate_mono.py
accelerate launch generate_stereo.py
To train Zero-RAFT-Stereo and Zero-IGEV-Stereo, run:
CUDA_VISIBLE_DEVICES='0,1' accelerate launch train_stereo.py
To evaluate Zero-RAFT-Stereo, run:
accelerate launch evaluate_stereo.py
To evaluate Zero-IGEV-Stereo, modify config/evaluate_stereo.yaml or run:
accelerate launch evaluate_stereo.py model=igev_stereo checkpoint=checkpoint/igev_stereo/model_700.safetensors
To save disparity, run:
accelerate launch save_disparity.py
This project is based on MfS-Stereo, Depth Anything V2, Marigold, RAFT-Stereo, and IGEV-Stereo. We thank the original authors for their excellent works.